mmdeploy/csrc/net/ort/ort_net.cpp

// Copyright (c) OpenMMLab. All rights reserved.

#include "ort_net.h"

#include <algorithm>

#include "core/logger.h"
#include "core/model.h"
#include "core/utils/formatter.h"
#include "onnxruntime_register.h"

namespace mmdeploy {

static TensorShape to_shape(const Ort::TypeInfo& info) {
  auto shape = info.GetTensorTypeAndShapeInfo().GetShape();
  return {shape.begin(), shape.end()};
}

static Result<DataType> ConvertElementType(ONNXTensorElementDataType type) {
  switch (type) {
    case ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT:
      return DataType::kFLOAT;
    case ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT16:
      return DataType::kHALF;
    case ONNX_TENSOR_ELEMENT_DATA_TYPE_INT8:
    case ONNX_TENSOR_ELEMENT_DATA_TYPE_BOOL:
      return DataType::kINT8;
    case ONNX_TENSOR_ELEMENT_DATA_TYPE_INT32:
      return DataType::kINT32;
    case ONNX_TENSOR_ELEMENT_DATA_TYPE_INT64:
      return DataType::kINT64;
    default:
      MMDEPLOY_ERROR("unsupported ONNXTensorElementDataType: {}", static_cast<int>(type));
      return Status(eNotSupported);
  }
}

// TODO: handle datatype
Result<void> OrtNet::Init(const Value& args) {
  auto& context = args["context"];
  device_ = context["device"].get<Device>();
  stream_ = context["stream"].get<Stream>();

  auto name = args["name"].get<std::string>();
  auto model = context["model"].get<Model>();

  OUTCOME_TRY(auto config, model.GetModelConfig(name));

  OUTCOME_TRY(auto onnx, model.ReadFile(config.net));

  Ort::SessionOptions options;
  options.SetLogSeverityLevel(3);

  RegisterCustomOps(options, OrtGetApiBase());

  if (device_.is_device()) {
    OrtCUDAProviderOptions cuda_options{};
    cuda_options.device_id = device_.device_id();
    // TODO set compute stream
    options.AppendExecutionProvider_CUDA(cuda_options);
  }
  session_ = {env_, onnx.data(), onnx.size(), options};

  auto memory_info = Ort::MemoryInfo::CreateCpu(OrtArenaAllocator, OrtMemTypeDefault);
  Ort::Allocator allocator(session_, memory_info);

  auto n_inputs = session_.GetInputCount();

  // force negative shape to be empty
  auto filter_shape = [](TensorShape& shape) {
    if (std::any_of(begin(shape), end(shape), [](auto x) { return x < 0; })) {
      shape = {};
    }
  };

  for (int i = 0; i < n_inputs; ++i) {
    auto input_name = session_.GetInputName(i, allocator);
    auto type_info = session_.GetInputTypeInfo(i);
    auto shape = to_shape(type_info);
    MMDEPLOY_INFO("input {}, shape = {}", i, shape);
    filter_shape(shape);
    OUTCOME_TRY(auto data_type,
                ConvertElementType(type_info.GetTensorTypeAndShapeInfo().GetElementType()));
    input_tensors_.emplace_back(TensorDesc{device_, data_type, shape, input_name});
    allocator.Free(input_name);
  }

  auto n_outputs = session_.GetOutputCount();

  for (int i = 0; i < n_outputs; ++i) {
    auto output_name = session_.GetOutputName(i, allocator);
    auto type_info = session_.GetOutputTypeInfo(i);
    auto shape = to_shape(type_info);
    MMDEPLOY_INFO("output {}, shape = {}", i, shape);
    filter_shape(shape);
    OUTCOME_TRY(auto data_type,
                ConvertElementType(type_info.GetTensorTypeAndShapeInfo().GetElementType()));
    output_tensors_.emplace_back(TensorDesc{device_, data_type, shape, output_name});
    allocator.Free(output_name);
  }

  return success();
}

Result<void> OrtNet::ForwardAsync(Event* event) { return Status(eNotSupported); }

Result<void> OrtNet::Deinit() { return success(); }

Result<Span<Tensor>> OrtNet::GetInputTensors() { return input_tensors_; }

Result<Span<Tensor>> OrtNet::GetOutputTensors() { return output_tensors_; }

Result<void> OrtNet::Reshape(Span<TensorShape> input_shapes) {
  for (size_t i = 0; i < input_shapes.size(); ++i) {
    input_tensors_[i].Reshape(input_shapes[i]);
  }
  return success();
}

static Ort::MemoryInfo MemoryInfo(const TensorDesc& desc) {
  const char* device_name = desc.device.is_host() ? "Cpu" : "Cuda";
  Ort::MemoryInfo memory_info(device_name, OrtDeviceAllocator, desc.device.device_id(),
                              OrtMemTypeDefault);
  return memory_info;
}

static Ort::Value AsOrtValue(Tensor& tensor) {
  auto memory_info = MemoryInfo(tensor.desc());
  std::vector<int64_t> shape(begin(tensor.shape()), end(tensor.shape()));
  return Ort::Value::CreateTensor(memory_info, tensor.data<float>(), tensor.size(), shape.data(),
                                  shape.size());
}

static Result<Tensor> AsTensor(Ort::Value& value, const Device& device) {
  auto info = value.GetTensorTypeAndShapeInfo();
  TensorDesc desc;
  desc.shape = info.GetShape();
  desc.device = device;
  OUTCOME_TRY(desc.data_type, ConvertElementType(info.GetElementType()));
  std::shared_ptr<void> data(const_cast<void*>(value.GetTensorData<void>()), [](void*) {});
  return Tensor(desc, data);
}

Result<void> OrtNet::Forward() {
  try {
    OUTCOME_TRY(stream_.Wait());
    Ort::IoBinding binding(session_);
    std::vector<Ort::Value> inputs;
    std::vector<Ort::Value> outputs;
    Ort::RunOptions options;

    inputs.reserve(input_tensors_.size());
    for (auto& t : input_tensors_) {
      inputs.push_back(AsOrtValue(t));
      binding.BindInput(t.name(), inputs.back());
    }

    // TODO: We are in the same situation as PPL.nn, the backend can't infer shapes
    //  without executing forward
    for (auto& t : output_tensors_) {
      binding.BindOutput(t.name(), MemoryInfo(t.desc()));
    }

    session_.Run({}, binding);

    outputs = binding.GetOutputValues();
    for (size_t i = 0; i < output_tensors_.size(); ++i) {
      OUTCOME_TRY(auto tmp, AsTensor(outputs[i], output_tensors_[i].device()));
      output_tensors_[i].Reshape(tmp.shape());
      OUTCOME_TRY(tmp.CopyTo(output_tensors_[i], stream_));
    }

    OUTCOME_TRY(stream_.Wait());
  } catch (const std::exception& e) {
    MMDEPLOY_ERROR(e.what());
    return Status(eFail);
  }
  return success();
}

class OrtNetCreator : public Creator<Net> {
 public:
  const char* GetName() const override { return "onnxruntime"; }
  int GetVersion() const override { return 0; }
  std::unique_ptr<Net> Create(const Value& args) override {
    try {
      auto p = std::make_unique<OrtNet>();
      if (auto r = p->Init(args)) {
        return p;
      } else {
        MMDEPLOY_ERROR("error creating OrtNet: {}", r.error().message().c_str());
        return nullptr;
      }
    } catch (const std::exception& e) {
      MMDEPLOY_ERROR("unhandled exception when creating ORTNet: {}", e.what());
      return nullptr;
    }
  }
};

REGISTER_MODULE(Net, OrtNetCreator);

}  // namespace mmdeploy